[Bug] LongCat-Image-Edit: crash (GPU+offload), mosaic output (CPU)

[Jared-D]

Git commit

commit c00a9e9, release master-778
Using prebuilt sd-master-c00a9e9-bin-Linux-Ubuntu-24.04-x86_64-rocm-7.2.1.zip

Operating System & Version

Fedora 44 Linux, x86_64

GGML backends

HIP, CPU

Command-line arguments used

~/stable-diff/sd-cli \
 --diffusion-model ./models/LongCat-Image-GGUF/LongCat-Image-Edit-Q8_0.gguf \
 --vae ./models/LongCat-Image-GGUF/vae/diffusion_pytorch_model.safetensors \
 --llm ./models/Qwen2.5-VL-7B-Instruct-GGUF/Qwen2.5-VL-7B-Instruct.Q8_0.gguf \
 --llm_vision ./models/Qwen2.5-VL-7B-Instruct-GGUF/Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf \
 --offload-to-cpu \
 --max-vram -1 \
 --diffusion-fa \
 -o ./output_edited_image-$(date '+%s').png \
 -p "Change to photo realistic style" \
 --ref-image ./city-512.png \
 --sampling-method euler \
 --flow-shift 3 \
 --cfg-scale 5.0 \
 --steps 20 \
 -v \
 -W 512 -H 512

Steps to reproduce

Run CLI as above.

Crash and stacktrace (full log available below):

[INFO ] conditioner.hpp:2360 - LongCatEditPipeline
[DEBUG] conditioner.hpp:2393 - resize conditioner ref image 0 from 512x512 to 504x504
/home/runner/work/stable-diffusion.cpp/stable-diffusion.cpp/ggml/src/ggml-impl.h:318: fatal error
/opt/working-dir/stable-diff/libggml-base.so.0(+0x1c1ab) [0x7f1af4ec81ab]
/opt/working-dir/stable-diff/libggml-base.so.0(ggml_print_backtrace+0x21c) [0x7f1af4ec862c]
/opt/working-dir/stable-diff/libggml-base.so.0(ggml_abort+0x15b) [0x7f1af4ec880b]
/opt/working-dir/stable-diff/libggml-base.so.0(+0x31364) [0x7f1af4edd364]
/opt/working-dir/stable-diff/libggml-base.so.0(ggml_gallocr_reserve_n_size+0x1b) [0x7f1af4eddf6b]
/opt/working-dir/stable-diff/libstable-diffusion.so(_ZN2sd14ggml_graph_cut30measure_segment_compute_bufferEP12ggml_backendP11ggml_cgraphRKNS0_7SegmentEPKc+0x266) [0x7f1af580fd66]
/opt/working-dir/stable-diff/libstable-diffusion.so(+0x18bb04) [0x7f1af5811b04]
/opt/working-dir/stable-diff/libstable-diffusion.so(_ZN2sd14ggml_graph_cut10build_planEP12ggml_backendP11ggml_cgraphRKSt13unordered_setIPK11ggml_tensorSt4hashIS8_ESt8equal_toIS8_ESaIS8_EEPKc+0x11a0) [0x7f1af5813f70]
/opt/working-dir/stable-diff/libstable-diffusion.so(_ZN2sd14ggml_graph_cut12resolve_planEP12ggml_backendP11ggml_cgraphPNS0_9PlanCacheEmRKSt13unordered_setIPK11ggml_tensorSt4hashISA_ESt8equal_toISA_ESaISA_EEPKc+0xfd) [0x7f1af581579d]
/opt/working-dir/stable-diff/libstable-diffusion.so(_ZN10GGMLRunner22resolve_graph_cut_planEP11ggml_cgraphPN2sd14ggml_graph_cut4PlanEPm+0xdc) [0x7f1af5839ecc]
/opt/working-dir/stable-diff/libstable-diffusion.so(_ZN10GGMLRunner7computeIfEESt8optionalIN2sd6TensorIT_EEESt8functionIFP11ggml_cgraphvEEibbbb+0x4c2) [0x7f1af5862d62]
/opt/working-dir/stable-diff/libstable-diffusion.so(_ZN3LLM9LLMRunner12encode_imageEiRKN2sd6TensorIfEEbbb+0x63) [0x7f1af5a94853]
/opt/working-dir/stable-diff/libstable-diffusion.so(_ZN11LLMEmbedder21get_learned_conditionEiRK17ConditionerParams+0x1aa2) [0x7f1af5abe312]
/opt/working-dir/stable-diff/libstable-diffusion.so(generate_image+0xffe) [0x7f1af591ffbe]
/opt/working-dir/stable-diff/sd-cli(+0x28b1a) [0x557e24be5b1a]
/lib64/libc.so.6(+0x3681) [0x7f1af4f89681]
/lib64/libc.so.6(__libc_start_main+0x88) [0x7f1af4f89798]
/opt/working-dir/stable-diff/sd-cli(+0x29235) [0x557e24be6235]

Reproduces identically (same stack trace) with both --diffusion-model as GGUF Q8_0 and as bf16 safetensors.
Adjusting the following parameters did not seem to fix the crash: without width or height specified, with or without --diffusion-fa, different --flow-shift, different --cfg-scale, different --steps.

For comparison the (non-edit) base LongCat-Image model generated a good text-to-image without mosaic noise (on GPU with --offload-to-cpu), sharing the same VAE and LLM files. So the crash seems specific to LongCat-Image-Edit, not LongCat-Image.

LongCat-Image text to image cli

``` ~/stable-diff/sd-cli \ --diffusion-model ./models/LongCat-Image-GGUF/LongCat-Image-BF16.gguf \ --vae ./models/LongCat-Image-GGUF/vae/diffusion_pytorch_model.safetensors \ --llm ./models/Qwen2.5-VL-7B-Instruct-GGUF/Qwen2.5-VL-7B-Instruct.Q8_0.gguf \ -o ./text2image-test-$(date '+%s').png \ -p "A photo of a city street" \ --sampling-method euler \ --flow-shift 3 \ --cfg-scale 5.0 \ --steps 20 \ -v \ -W 1024 -H 1024 \ --max-vram -1 \ --offload-to-cpu ```

What you expected to happen

To run, not crash, and produce an edited output image.

What actually happened

A crash.

Logs / error messages / stack trace

LongCat-Image-Edit-with-GPU.log

Additional context / environment details

Workaround: --backend cpu avoids the crash.

Secondary issue: mosaic output on --backend cpu

With the crash avoided via --backend cpu, output shows the reference image's overall shape but is covered in mosaic/scrambled block artifacts. A few different image resolutions were tried but with the same scrambled problem.

Command:

~/stable-diff/sd-cli \
 --diffusion-model ./models/LongCat-Image-GGUF/LongCat-Image-Edit-Q8_0.gguf \
 --vae ./models/LongCat-Image-GGUF/vae/diffusion_pytorch_model.safetensors \
 --llm ./models/Qwen2.5-VL-7B-Instruct-GGUF/Qwen2.5-VL-7B-Instruct.Q8_0.gguf \
 --llm_vision ./models/Qwen2.5-VL-7B-Instruct-GGUF/Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf \
 -o ./out.png \
 -p "Change to photo realistic style" \
 --ref-image ./city-512.png \
 --sampling-method euler \
 --flow-shift 3 \
 --cfg-scale 5.0 \
 --steps 5 \
 -v \
 -W 512 -H 512 \
 --backend cpu

Log:
cpu-run.log

Output image:

Reference image:

Running the same sort of edit task using a different (FLUX.1-Kontext-dev) "edit" model does work successfully.

FLUX.1-Kontext-dev image edit cli

``` ~/stable-diff/sd-cli \ --diffusion-model ./models/FLUX.1-Kontext-dev-GGUF/flux1-kontext-dev-Q5_K_M.gguf \ --vae ./models/FLUX.1-dev/ae.safetensors \ --clip_l ./models/flux_text_encoders/clip_l.safetensors \ --t5xxl ./models/flux_text_encoders/t5xxl_fp16.safetensors \ --offload-to-cpu \ --max-vram -1 \ --diffusion-fa \ -o ./output_edited_image-$(date '+%s').png \ -p "Change to photo realistic style" \ --ref-image ./city-512.png \ --sampling-method euler \ --clip-on-cpu \ --cfg-scale 1.0 \ -v \ --steps 20 ```